Evolving Explanatory Novel Patterns for Semantically-Based Text Mining - Natural Language Processing and Text Mining - page 164

• THEN, in vitro digestion of three cutting ages in six ecotypes has an effect on

”Bigalta” cuttings which got their higher performance in a 63-day period.

This hypothesis looks more complete (goal, methods, etc.) but is less relevant
than the previous hypothesis despite its close coherence. Note also that the plausi-
bility is much higher than for hypothesis 65, but the other criteria seemed to be a
key factor for the experts.
The hypothesis concerns the production and cutting of a specific kind of tree (Pi-
nus) and forests where these lie. However, the second role (“the method is based...”)
discusses a different topic (mice capture) which apparently has nothing to do with
the main issue and that is the reason for the pattern's coherence to be scored lower
than the previous hypothesis (41% vs. 50%). The model also discovered that there
are organisms (and issues related to them) which are affecting the Pinus (and for-
est) restoration (i.e., mice). This fact has received a higher value for Plausibility of
Origin or the Novelty of the pattern (99%) and consequently, it is correlated with
the experts opinion of the pattern (score=3.20).
Another example of discovered patterns is a low-scored hypothesis given by the
following hypothesis 52:

IF object(perform(20611)) AND
object(carry_out(2631))
THEN effect(1931,1932)

and has a criteria vector [ 0 . 29 , 0 . 48 , 0 . 49 , 0 . 014 , 0 . 2 , 0 , 0 . 3 , 0 . 5] and obtained an av-
erage expert's assessment of 1 . 53.
The structure of this pattern (48%) is better than for hypothesis 88. However,
since the hypothesis is not complete, this has been scored lower than the previous
one. This might be explained because the difference in structure between object-
object and goal-method (Figure 9.3) is not significant and as both hypotheses (88
and 52) become final solutions, the expert scored best those which better explain
the facts. Note that as the model relies on the training data, this does not ensure
that every hypothesis is complete. In fact, previous experimental analyses of recall
show that only 26% of the original rules representing the documents contain some
sort of “method.”
In natural-language text, the pattern can roughly be interpreted as:

• IF the object of the work is to perform the analysis of the fractioned honey

in Brazil for improving the producers' income and profitability, AND

• The object of the work is to carry out observations for the study of Pinus

hartwegii at the mexican snowed hills so to complement the previously existing
information about the development status of Adjunctus and its biology.

• THEN in vitro digestion of three cutting ages in six ecotypes has an effect on

bigalta cuttings which got their higher performance in a 63-day period.

This hypothesis shows the same relevance as the previous one (29%) indicating
that both attempt to explain the connection between the target concepts and that
contained in the pattern. Note also that coherence is not very high (49%) consider-
ing that one part of the pattern discusses the “honey production” and issues, and
the other parts deal with the investigation, production and cutting of Pinus trees.
Accordingly the degree of interestingness and novelty of the patterns has been low